How AI Engines Choose Which Sources to Cite

AI engines choose which sources to cite by running a rough four-stage pipeline: crawl access, retrieval of candidate pages, synthesis of an answer, and a final citation-selection step where most of the retrieved pages get dropped. The punchline that trips everyone up: getting retrieved is not getting cited. According to AirOps, only 15 percent of 548,534 retrieved pages actually made it into a final answer. The other 85 percent got read, judged, and tossed.

So the question is not "can the engine find me." That's table stakes. The question is whether your page survives all four stages, especially the last one, where the model has a pile of candidate sources open and picks which handful to quote and credit. This is the technical mental model for that whole machine: what each stage does, what you actually control at each step, and where most pages quietly bleed out.

I'm going to be blunt about the parts that are oversold. Some of this is genuinely under your control. Some of it is just being indexed and ranking like it's 2015. And some of it is luck plus cross-source agreement that you can nudge but not force. Let's walk the pipeline.

Key takeaways

• Retrieval is not citation. According to AirOps, only 15 percent of 548,534 retrieved pages got cited. The final selection step throws away 85 percent of what the engine reads.
• Fan-out is the norm. AirOps found fan-out (the engine firing multiple sub-queries) happened on 89.6 percent of searches, and 32.9 percent of cited pages were reachable only through fan-out, not the original query.
• Classic ranking still feeds the funnel. AirOps found 55.8 percent of cited pages also ranked in Google's top 20, and a page at position 1 had a 43.2 percent citation rate. Old-school SEO gets you into the candidate set.
• The top of the page does the work. According to Omnibound, 44.2 percent of AI citations come from the first 30 percent of the content. Front-load the answer or lose it.
• Engines behave very differently. Onely measured citation rates of 13.05 percent on Perplexity, 6.38 percent on Gemini, 2.11 percent on Google AI Overviews, and 0.59 percent on ChatGPT. One strategy does not fit all four.

The four-stage pipeline in one breath

Here is the whole machine before we zoom into the parts. An AI engine answers a question by, first, having already crawled and indexed pages it's allowed to access. Second, when a query comes in, it retrieves a candidate set of pages it thinks are relevant, often by firing several sub-queries instead of one. Third, it synthesizes an answer from those candidates, reading and weighing them. Fourth, it selects which of those candidates to cite, and most of them don't make the cut.

Access, then retrieve, then synthesize, then cite. Every stage is a filter. You can be perfect at stages one through three and still lose at stage four, because the model read your page, decided a competitor said it better, and credited them instead. That's the part that feels unfair until you understand it. (It's not unfair. It's just ranking with extra steps.)

The useful mental model is a funnel that narrows hard at the end. Lots of pages are eligible. Fewer get retrieved. Fewer get read closely. A tiny slice gets cited. Each section below is one narrowing point, and your job is to figure out which one is killing you.

Stage 1: access, where most invisibility actually starts

Access is the boring stage nobody wants to hear about, and it's where the largest share of "why am I invisible" problems live. If an AI engine's crawler cannot fetch and render your page, nothing downstream matters. You're not in the index, so you're not in the candidate set, so you cannot be retrieved, so you cannot be cited. Game over before the game starts.

Four things commonly block access. First, robots.txt rules that disallow the AI crawler, often by accident. Loads of sites blocked GPTBot back in 2023 to keep their content out of model training, then used a blanket rule that also caught the search crawler, and locked themselves out of ChatGPT search without realizing it. OpenAI is clear in its bot documentation that sites opted out of OAI-SearchBot will not appear in ChatGPT search answers. So check your file. We wrote a whole breakdown of which bots to allow in the guide to AI crawlers and robots.txt.

Second, your WAF or CDN bot rules. Cloudflare, Akamai, and friends ship "block AI bots" toggles, and a security team can flip one without telling marketing. The crawler gets a 403, you never find out, and your logs are the only place the truth shows up. Third, JavaScript rendering. If your content only appears after client-side JS runs, and the crawler doesn't execute that JS, it sees an empty shell. Server-side render or pre-render the stuff you want quoted. Fourth, the newer llms.txt convention, which is a guide file some sites publish to point engines at their best content. It's optional and support is uneven, so don't oversell it, but it costs little. We cover what it does and doesn't do in the llms.txt explainer.

The honest summary: access is the one stage you fully control, and it's the cheapest to fix. Check robots.txt, check your WAF logs, check that your content exists in raw HTML. Most invisibility starts right here.

Stage 2: retrieval, the candidate set and fan-out

Retrieval is where the engine, given a live query, pulls a set of candidate pages to read. This is not one search. It's usually several. AirOps found that ChatGPT fans out, firing multiple related sub-queries, on 89.6 percent of searches. Ask it "best CRM for a small team" and it might quietly search "CRM pricing for startups," "easiest CRM to set up," and "CRM with email integration" all at once, then pool the results.

That fan-out behavior matters more than people expect. AirOps found that 32.9 percent of cited pages were reachable only through a fan-out sub-query, not the user's original phrasing. In plain terms: a third of cited pages would never have been found if the engine had searched only the literal question. So you're not optimizing for one keyword. You're trying to be the obvious answer to a cluster of related questions around your topic. That's an entity and topical-coverage game, which is why entity SEO keeps coming up in this space.

Where does the candidate set come from? Mostly a search index. ChatGPT's web search leans heavily on the Bing index, per OpenAI's own bot docs, and the others ground on Google or their own crawls. This is why traditional ranking still feeds the funnel hard. AirOps found 55.8 percent of cited pages also ranked in Google's top 20, and a page sitting at Google position 1 had a 43.2 percent citation rate. Freshness matters too, since grounding systems favor recently updated pages for queries that look time-sensitive.

So retrieval rewards two things at once: classic SEO fundamentals that get you into the index and ranking well, plus broad topical coverage so you match the fan-out sub-queries. Neither alone is enough. If you want the full strategic stack on this, the generative engine optimization guide lays it out end to end.

Stage 3: synthesis, where the model writes the answer

Synthesis is the stage where the model actually reads the candidate pages and writes a response. This is the part people imagine when they think "AI search," but it's only one filter of four. The model has a stack of retrieved pages open, it drafts an answer from them, and it decides which claims to make and roughly where they came from. The page content gets weighted unevenly during this read.

The big practical fact: the early part of a page carries disproportionate weight. According to Omnibound, 44.2 percent of AI citations come from the first 30 percent of the content. Models read top-down, context windows have limits, and the opening is where a self-contained answer is most likely to sit. If your direct answer is buried under 600 words of throat-clearing about how AI is changing everything, the model may synthesize its answer before it ever reaches your good part.

So synthesis rewards answer-first writing in clean, liftable chunks. Short paragraphs. One idea per section. A plain definition sentence the model can grab whole ("X is ..."). Tables it can read row by row. This is craft-level work on the actual sentences, and it's the most controllable lever after access. We get specific about the writing itself in the AI content optimization guide. The mental model: make the easiest correct thing for the machine to quote, and put it near the top.

Stage 4: citation selection, the brutal final filter

Citation selection is the stage that decides everything and the one almost nobody optimizes for directly. After synthesis, the engine picks which of the candidate pages to actually attribute. Here is the number that should reset your strategy: AirOps found only 15 percent of those 548,534 retrieved pages got cited. The model read 85 percent of them, used them or didn't, and chose not to credit them. They were good enough to retrieve and not good enough to cite.

What survives this filter? Three things show up over and over. First, fact density: pages packed with concrete, checkable facts, numbers, and named sources give the engine specific things to attribute, and a citation is basically the model saying "I got this exact fact here." Vague pages have nothing worth crediting. Second, quotability: a clean, self-contained sentence or row the engine can lift without rewriting your whole paragraph. Third, agreement: when your claim lines up with what other retrieved sources say, the engine trusts it more and is likelier to cite the source that states it cleanest.

That last one is the quiet kingmaker. If five sources agree on a fact and your page states it most precisely with a named source attached, you win the citation even against bigger brands. If your page is the lone outlier making a claim nobody else supports, the model tends to route around you, because citing a contradicted source is risky for the engine. This is the difference between a brand mention and a real citation, which we untangle in the citation rate glossary entry.

The takeaway is uncomfortable and freeing at the same time. You cannot bully your way to a citation with brand size or backlinks alone. You earn it by being the clearest, most factual, most agreeable-with-the-consensus statement of the answer on the page. That's a writing and structure problem, and it's winnable by small sites.

The whole pipeline as one table

Here's the funnel in one view: what the engine does at each stage, and the lever you actually control. Pin this somewhere.

Read it left to right and you can diagnose almost any visibility problem. No crawl access means nothing else fires. Good access but no retrieval means a ranking and coverage problem. Retrieved but never cited means a stage-four problem: your content got read and judged not worth quoting. Different stage, different fix. Treating "I'm not in ChatGPT" as one undifferentiated problem is why so many GEO efforts spin their wheels.

Why agreement and entities decide the close calls

Two forces tip the final selection, and both reward the same behavior. The first is cross-source agreement, which I described above. AI engines are nervous about citing claims that other retrieved sources contradict, because a wrong citation makes the engine look bad. So when your fact matches the consensus and you state it crisply with a named source, you become the safe pick. Disagreeing with the field is a fine editorial stance for a human reader and a quiet way to lose citations.

The second force is entities. An entity is a thing the engine recognizes as a distinct, known item: a company, a person, a product, a concept, with a stable identity across the web. When your brand is a well-formed entity, with consistent naming, a clear "what it is" definition, and corroboration across sources, the engine can confidently attach a citation to you. When your brand is a fuzzy blob the model isn't sure about, it hesitates. This is where schema markup for AI search earns its keep, because structured data spells out who you are and what each page answers in a format the machine doesn't have to guess at.

Put those together and the close-call rule is simple. The engine cites the source that states the consensus answer most clearly and comes from the most clearly defined entity. Be precise, be corroborated, be recognizable. Boring advice, but it's what actually moves the needle.

How this differs across ChatGPT, Perplexity, Gemini, and AIO

Same web, four very different machines. The pipeline shape is roughly shared, but each engine has its own crawler, candidate budget, grounding source, and appetite for citing at all. The cleanest evidence is the raw citation rate per engine. Onely measured how often each engine cites external sources, and the spread is wild.

Source: Onely's 2026 brand visibility research.

Read that and a few things click. Perplexity cites constantly, so it's the friendliest place to earn visibility and a sane place to start testing. ChatGPT cites rarely per response, which makes each citation precious and the stage-four filter especially brutal, so answer-first precision matters most there. The practical guides diverge accordingly: how to get cited by ChatGPT is heavy on Bing indexing and robots.txt, while how to rank on Perplexity leans on freshness and tight factual chunks. Gemini and Google AI Overviews both ride Google grounding, so strong Google ranking does double duty, which is the throughline in how to appear in Google AI Overviews.

The mistake is assuming one number describes "AI search." It doesn't. A page that crushes on Perplexity can be invisible in ChatGPT for reasons that have nothing to do with quality and everything to do with which index each engine grounds on. Measure them separately or you'll draw the wrong conclusion.

How to debug where you're losing

You can find the leaking stage in about an afternoon if you go in order. Don't rewrite your whole site before you know which filter is dropping you.

Step 1: confirm access

Pull your server logs and look for the AI crawler user agents: OAI-SearchBot, PerplexityBot, Google-Extended, and friends. Are they hitting your pages and getting 200s? If you see 403s or no visits at all, you have an access problem, and no content change will help until you fix robots.txt and your WAF rules. Then confirm your target content exists in raw HTML, not just after JavaScript runs.

Step 2: confirm retrieval and ranking

Check whether you rank in Google's top 20 and are indexed in Bing for your target queries and their obvious variants, since retrieval grounds on those indexes and fan-out hits the variants. If you're not indexed or ranking nowhere, that's a stage-two problem, and it's a classic SEO fix wearing a GEO hat. Broaden topical coverage so you match the fan-out sub-queries, not just the head term.

Step 3: confirm citation, then measure it over time

This is the one you cannot eyeball reliably. Citation behavior is probabilistic and varies by phrasing, engine, and day, so a single manual check tells you almost nothing. You need to run the same prompts repeatedly across engines and track how often you actually get cited, with a confidence interval, against your competitors. That's exactly what AI Citation Monitor is built for: it measures citation rate and competitor share of voice per engine over time and points at the specific fix. You can also run a free instant check to see where you stand right now. For the broader monitoring picture, AI citation tracking goes deeper.

The discipline that separates people who improve from people who guess: change one thing, then measure the citation rate before and after. Unblock the crawler, measure. Rewrite the intro answer-first, measure. Add the table and the named stats, measure. Without the measurement loop you're just redecorating and hoping. If your brand is flat-out missing, why your brand is not showing up in ChatGPT walks the same funnel from the symptom side.

The short version

AI engines choose sources through access, retrieval, synthesis, and citation selection, and each stage is a filter that drops pages. Most invisibility starts at access, most retrieval rides on classic ranking plus fan-out coverage, synthesis rewards answer-first chunks, and citation selection keeps only about 15 percent of what gets read. Win the last stage by being the clearest, most factual, most consensus-aligned, most recognizable statement of the answer. Then measure it per engine, because the four behave nothing alike. That's the whole machine.

FAQ

How do AI engines decide which sources to cite?

AI engines run a four-stage pipeline: crawl access, retrieval of candidate pages, synthesis of an answer, and a final citation-selection step. Most retrieved pages never get cited. According to AirOps, only 15 percent of 548,534 retrieved pages were cited in the final answer. The pages that survive tend to be the ones that directly answer the query, pack real facts near the top, and agree with what other sources say.

Is getting retrieved the same as getting cited?

No, and confusing the two is the most common mistake in GEO. Retrieval just means the engine pulled your page into its candidate set to read it. Citation means the engine actually used and credited you in the answer. AirOps found 85 percent of retrieved pages get read, judged, and dropped, so retrieval is necessary but nowhere near sufficient.

Does my Google or Bing ranking still matter for AI citations?

Yes, a lot more than people assume. AirOps found 55.8 percent of cited pages also ranked in Google's top 20, and a page ranking first in Google had a 43.2 percent chance of being cited. ChatGPT's web search leans on the Bing index, so classic indexing and ranking still feed the candidate set. Strong traditional SEO does not get you cited by itself, but it gets you into the room.

Why does the start of my page matter so much?

Because synthesis weighs the early part of a page heavily, and most citations come from there. According to Omnibound, 44.2 percent of AI citations come from the first 30 percent of the content. Put your direct, self-contained answer in the first two or three sentences, before any setup, so the part of the page the engine reads first is the part worth quoting.

Why do citation rates differ so much between engines?

Because each engine has its own retrieval system, candidate budget, and appetite for citing. Onely measured citation rates of 13.05 percent on Perplexity, 6.38 percent on Gemini, 2.11 percent on Google AI Overviews, and 0.59 percent on ChatGPT. Same web, very different behavior, which is why you measure each engine separately instead of assuming one number covers all of them.

How do I find out where I am losing in the pipeline?

Debug one stage at a time. Check server logs and robots.txt for access, check whether you rank and get retrieved for the target queries, then check whether retrieved pages actually get cited. A tool like AI Citation Monitor tracks citation rate and share of voice per engine over time so you can see which stage is leaking instead of guessing.